The Crown Joules: What AI Actually Costs to Run
RishiDean_AIEnergyUsage
Share

The Crown Joules: What AI Actually Costs to Run

The narrative that every AI query is melting the planet makes a great soundbite, but is bad math. Here's a technical breakdown of AI energy usage.

The narrative that every AI query is melting the planet makes a great soundbite, but is bad math. Here's a technical breakdown of AI energy usage.

I keep hearing some version of this: every time you ask ChatGPT for a recipe, you're burning a rainforest (or some other variant). Trust me, I'm the last person to speak to the morality of energy use. But I am someone who can explain how these systems actually work from a technical perspective. We carefully manage context windows, and token usage, and get billed every month, so the math is highly measurable.

Let me break it down.

How AI Uses Energy

There are two types of energy consumption in AI, and conflating them is where most of the confusion starts.

Training is when companies like Anthropic, OpenAI, or Google build a model from scratch. These are massive, resource-intensive runs. Generally the big foundational model releases are relatively infrequent (a few times a year), but they, and others, are being trained and released often. Training is the energy-hungry part of AI, which we'll talk more about in a second.

Inference is what happens every time we ask a question. The model is already built, so we're just using it, and this is orders of magnitude more efficient.

Over the lifetime of a model, inference is where a majority of AI's TOTAL energy consumption lives (since the training cost gets amortized over the life of the model). Some estimates are as high as 90% of a model's lifetime energy use.

Tokens: the fundamental AI unit

AI models don't process words the way we think about them. Under the hood, they operate on something called tokens, which you can think of as the fundamental unit of AI computation.

A token isn't exactly a word. It's more like a chunk of text the model operates on. Roughly, 1 word ≈ 1.3 tokens but the relationship isn't totally correlated. The reason this matters is that energy consumption scales with tokens processed, not words typed. But yes, more words = more tokens = more energy.

Because, we, as users, are actively tracking and monitoring each session, we can readily report on our usage stats, and translate that to dollars.

Mapping tokens to energy use

It does vary with model, hardware, and other factors, but a single token costs anywhere from about 0.4 to 4 Joules to process. For context, lifting an apple one meter off a table takes about 1 Joule.

But this is also where efficiency is improving at breakneck pace. Google recently published that over just 12 months, the energy cost of a median Gemini query dropped by 33x. Across the industry, while precise multipliers are hard to verify, energy efficiency per token has improved dramatically since the early GPT-3 era. This technology is getting dramatically cheaper to run, not more expensive.

Understanding AI energy use vs. daily life

Let's convert this into units your electric company would actually bill you:

So, if you blended a smoothie this morning, or charged your car, or used a hairdryer, you've already blown past that for the day…before even leaving the house.

The Bigger Picture

All U.S. data centers for everything — not just AI, but Netflix, your bank, FaceTime calls, scrolling TikTok, and every service you've ever been on — used about 176 terawatt-hours in 2023. That's roughly 4.4% of total U.S. electricity. And remember, AI is a subset of that number.

And it's not just energy, but water usage also tells a similar story. In Maricopa County, Arizona, golf courses use 30x more water than data centers, while generating 50x less tax revenue per unit of water consumed, and probably only used by a couple hundred rich dudes.

Nationally, data centers account for about 0.15–0.19% of U.S. freshwater consumption. Golf courses across the country collectively use over a billion gallons of water per day. But also think about casinos, fast food production, almond milk, and everything else that uses much more water.

What This Means

I'm not trying to say that the planet doesn't matter. It absolutely does. But AI isn't the boogeyman many think it is, and the efficiency trajectory is heading in the right direction, and at a pace that's unusual for any technology. If we're serious about helping the planet, there are far less utility producing things to society to start with….ahem, golf courses.

So while folks on TikTok rant, the data doesn't lie.